Challenge: The client, a multinational pharmaceutical company wanted to refactor an existing system for enquiry management and improve its performance and analytical capabilities through AI. The goal was to ensure a consistent customer experience across multiple locations. 

Industry: Healthcare  

Approach: Our team used semantic analysis with NLTK preprocessing to create different feature extraction and drug product labelling over already accumulated and classified еnquiry data. Preprocessed data were used to train Keras with the TF engine, and evaluated versus built-in SVM algorithms in SKLearn. Results were accuracy near 80% for classification. ML components were integrated into the auto-assign pipeline based on product and team recognized in free text еnquiry.

From a business perspective, the complete solution speeds up the process of enquiry management and contributes to overall customer satisfaction and increased operational efficiency. 

Tech stack: Python and Angular developers, Data Scientists, ML Engineers, Project Manager.

Overview: The client was a start-up solving the problem of managing numerous processes within a recruitment department. Hiring teams have data from a variety of sources, which are difficult to process and analyze. Our task was to structure and develop a data structure and create an AI tool for screening candidates.

Approach: Our team of data engineers, data scientists, and ML engineers started with a research phase to understand the client's requirements and identify the best approach for the project. We conducted a thorough analysis of the recruitment data landscape and the different types of data sources used in the industry. We then developed a solution architecture that could handle the complex data structures and AI models required for the project. We selected AWS as the primary provider for data storage and combined it with other ETL tools to work with different formats from multiple sources.

Once the data structure and pipelines were set up, we used semantic analysis with NLTK preprocessing to extract the necessary information for each candidate. The semantic analysis allowed us to identify the candidate's skills, experience, and education from their resume, cover letter, and social media profiles. We then developed the ML model using Keras, TensorFlow, and SKLearn to sort the candidates into predefined groups and factors, such as experience level, job title, and location. The output was presented in a structured database of candidates, which could be easily searched and filtered by hiring teams.

Challenge: The task was to develop a solution for task management in production companies combined with extracting data from production machines through sensors and PLCs.

Team: Python developers, Angular developers, Data scientists, Computer vision experts, QA experts, DevOps, UX designer and PM

Solution: The solution was successfully developed for the MVP phase and continued to include computer vision combined with tasks, procedures and read-only registers from PLC controllers. Currently using Python OpenCV & YOLO object detector integration from IP camera stream to extract workers' movements around specific industrial areas. We collect data first to perform supervised learning in later stages.

Status: Currently, the software is being integrated into 3 production plants in Bulgaria.

Challenge: Creating a fully functioning solution from scratch

Team: Data Engineers, Data Scientists, ML Engineers, DevOps

Solution: The project started with the investigation phase to analyse the existing data and test multiple solutions which meet the needs of both client and cloud providers. AWS was selected as a core service provider for data storage combined with ElasticSearch. Our team built the full ETL process to start operations. Once the data structure and pipelines were set, we introduced two teams of Data Scientists - one specialized in semantic analysis (NLP) and the other with experience in image recognition. The ML part was further taken care of by our engineers. The output was a dashboard of data visualisations made in Kibana to present the results.

Status: Successful completion of the data engineering phase with ongoing data science stages.

Challenge: The client`s request involved road and traffic analysis. The task included building a vehicle counting system based on vehicle type and classification, as well as differentiating between various types of vehicles and conducting a total count.

Approach: We started by developing a custom video surveillance software which made it possible to remotely watch the traffic situation and collect data. Utilizing image recognition and data analytics techniques we developed a POC model analysing the traffic data, which was further extended to a fully functioning tool for road and safety as the traffic analysis.

Through advanced deep learning our team created a solution which was able to detect and alert for potential accidents, predict congestion, or plan efficient roads and parking spaces. The solution is currently growing into autonomously monitoring traffic cameras and systems and generating real-time alerts when certain events of interest occur.

Tech stack: Python, React.JS, Yolo, SSD and OpenCV

Overview: Our client is a European MedTech start-up focused on the detection of cardiovascular anomalies. The start-up’s objective is to challenge the market for ECG signal-based detection of heart pathologies, such as, but not limited to, atrial fibrillation, premature beats, subventricular tachycardia, etc, obtained from affordable wearable devices such as the Polar H10. The company required an ML-based model development and preparation for the Class 2A medical device certification, in order to ensure that their system is safe and effective for patients to use at home. 

Team: Our team of experts included a Data Engineer, Data Scientists, ML Engineers and a Project Manager

Approach: 
Discovery phase

Our Data Scientists collaborated with the client in order to understand their existing ECG signal processing system, as well as their computation-based algorithms. We created an automated testing tool that assessed the performance of the current solution against a range of internal and external medical data sets. Based on this analysis, we provided recommendations to improve the scope of detectable medical conditions and identified areas for improvement. 
We advised the client on the certification process and prepared an action plan for the  ML-related work and documentation needed for the Class 2A certification. We also provided improvement recommendations and execution support, which helped the client enhance their ECG signal processing system and algorithms. Our team completed the analysis and recommendations in less than 8 weeks, providing actionable advice, which helped the client reduce the application processing time. 

ML-based signal processing model development phase

The client provided the raw ECG signal data, including labelled data with annotations of R-peaks, pathologies, as well as ‘soft’ and ‘hard’ noise segments. This labelled data was used as inputs for the models and deemed as their training set. The quality and accuracy of pathology predictions (detection) heavily depend on the cleaness of the ECG signal. This is why it is important to have robust, well-performing models in terms of accuracy and precision, which can distinguish between the clean and noisy parts of the ECG signals. Our Data Scientists developed ML-based models for ‘soft’  and ‘hard’ noise detection. Upon successfully passing the strict model performance acceptance criteria of the client, our ML Engineers created the production pipelines for training and inference and integrated it into the client’s platform and business processes. Our collaboration continues with the development of further pathology detection ML-based algorithms. 

Overview: A retail bank faced the challenge of optimizing offers to attract new clients and upsell existing clients. The marketing and risk departments were working with different approaches, and there was a need for a unified score that could provide a single measure of expected profitability per client. To address this challenge, our data science team was tasked with developing an AI solution that produces a robust objective score for the expected profitability per client, based on both marketing and risk factors.

Solution: Our team developed a CRM (Customer Relationship Management) model that combines marketing and risk factors to evaluate the expected profitability of each banking product based on our developed methodology product. The model uses Propensity-to-Buy (PtB) models to predict the likelihood of a customer purchasing a specific product, and Probability of Default (PD) models to estimate the risk of the customer defaulting on a loan. The output of the model is a single measure of expected profitability per client, which can be used to rank the Next-Best-Offer (NBO) for each customer. 

To develop the CRM model, we first collected and integrated data from various sources, including transactional data, demographic data, and customer interaction data. We then applied various machine learning algorithms, such as logistic regression and random forests, to train and test the PtB and PD models. We also conducted feature engineering and selection to identify the most relevant variables for predicting profitability.

Overview: The banking industry faces a significant challenge in accurately assessing the creditworthiness of borrowers to make informed lending decisions. Traditional credit scoring models, which often rely on limited information such as credit history, may not accurately reflect the risk of default. To address this challenge, our team was tasked with developing a more accurate and comprehensive model for predicting the probability of default.

Solution: Our team developed a PD (Probability of Default) model that uses a wide range of customer data, including behavioural and demographic data, to predict the probability of default. The model is based on machine learning algorithms that are trained on historical data to identify patterns and correlations between different variables and default risk. 

We first collected and cleaned a large dataset of loan applications and associated customer data and used various machine learning algorithms, such as logistic regression and decision trees, to train and test the model. We also conducted feature selection and engineering to identify the most relevant variables for predicting default risk.

The output of the model is a single score that reflects the probability of default, which can be used to inform lending decisions and set appropriate interest rates.

Challenge: Between 2009 and 2013, our Head of Data Science led the AI development and implementation of the euroCat project, a groundbreaking solution for personalized treatment methodology in radiation oncology. The project involved five hospitals across three EU countries, utilizing the data of their cancer patients with the aim to optimize personalized treatments. The goal was to develop a Machine Learning model that is capable of learning from disparate data sources without centralizing patient data, in a way that preserves patient privacy. You can read more about euroCat in the academic paper.

Solution: The solution leveraged Federated Learning (which has become the more common name for distributed learning) - a technique that enables multiple parties to collaborate on a Machine Learning model without sharing sensitive data and without the need for data to be centralized, via exchanging only model parameters during the learning process.
In this approach, each hospital trains part of the model at each iteration using its local data and then shares only some parameters with a central server. The central server then aggregates the parameters from all countries to update the global model without accessing the local data. The learning continues through iteration cycles until model convergence criteria are met. The role of a “central server” can be played by any of the hospitals on the grid. Georgi Nalbantov’s AI team used Federated Learning to learn Support Vector Machine (SVM) models, using the Alternating Direction Method of Multipliers (ADMM), from disparate databases to predict treatment outcomes: which can be either a direct treatment effect or a treatment side effect, for example, shortness of breath after (lung) radiotherapy.

The performance of the SVM models was evaluated by the Area Under the Curve (AUC) in a five-fold cross-validation procedure (training at four sites and validation at the fifth). The performance of the pooled (federated) learning algorithm was compared with centralized learning, in which the datasets of all clinics are combined in a single dataset. The result of the centralized model was (naturally) the same as the federated-learning model, as they are mathematically yielding the same result.